🤖 AI Summary
Addressing the challenge of balancing term accuracy and contextual adaptability in multilingual machine translation, this paper proposes DuTerm—a dual-stage architecture. The first stage employs a terminology-aware neural machine translation (NMT) model to generate an initial translation; the second stage leverages a large language model (LLM) as a context-driven post-editor—not a generator—to refine terminology consistency via prompt engineering. Crucially, the LLM is designed as a lightweight, controllable term calibration module, mitigating fluency degradation caused by excessive constraints. Experiments on the WMT 2025 Terminology Sharing Task (English–German/Spanish/Russian) demonstrate that DuTerm significantly outperforms strong baselines in both BLEU and term adherence metrics, with particularly notable improvements in term consistency under complex contextual conditions. These results validate the effectiveness and generalizability of the “NMT + LLM correction” paradigm.
📝 Abstract
This paper introduces DuTerm, a novel two-stage architecture for terminology-constrained machine translation. Our system combines a terminology-aware NMT model, adapted via fine-tuning on large-scale synthetic data, with a prompt-based LLM for post-editing. The LLM stage refines NMT output and enforces terminology adherence. We evaluate DuTerm on English-to German, English-to-Spanish, and English-to-Russian with the WMT 2025 Terminology Shared Task corpus. We demonstrate that flexible, context-driven terminology handling by the LLM consistently yields higher quality translations than strict constraint enforcement. Our results highlight a critical trade-off, revealing that an LLM's work best for high-quality translation as context-driven mutators rather than generators.